“Glow: Generative Flow with Invertible 1x1 Convolutions,” posted to arXiv on July 9, 2018 by Diederik P. Kingma and Prafulla Dhariwal of OpenAI, advanced a third family of deep generative models known as normalizing flows. Unlike GANs and VAEs, a flow-based model is built entirely from invertible transformations, which lets it compute the exact probability it assigns to any data point rather than only an approximation.
Glow’s main contribution was a learnable, invertible one-by-one convolution that generalized the fixed channel-shuffling operations used in earlier flow models. Combined with the rest of its architecture, this let Glow achieve strong likelihood scores on standard image benchmarks. More visibly, the authors showed that a model trained purely to maximize likelihood could nonetheless generate realistic, high-resolution faces and support smooth, semantically meaningful edits, such as changing a person’s expression or hair, by manipulating the latent representation.
Flow-based models like Glow never matched GANs or diffusion for raw image quality, but their exact-likelihood property kept them important for tasks where calibrated probabilities matter, and the mathematics of invertible transformations fed into later work on continuous normalizing flows and flow matching. For a general reader, Glow rounds out the picture of how researchers explored several distinct routes to the same goal of teaching machines to generate convincing images, each with its own trade-offs.